RNA 3D structure evaluator trained to rank candidate conformations of one RNA against each other and pick out the near-native ones.
No providers recorded yet. Browse all providers
An RNA structure predictor typically returns hundreds or thousands of conformations for a single sequence, and someone has to decide which one to believe. A structural biologist doing that job compares the candidates for that one molecule with each other rather than assigning each an absolute error in isolation. Learned scoring functions for RNA have mostly done the opposite: they regress the RMSD of every candidate to its native structure, with mini-batches mixing candidates from different RNAs whose RMSD ranges differ widely, so long molecules with broad error distributions dominate the loss and the relative differences between near-identical candidates are never directly supervised.
HomRank — homogeneous RNA ranking — trains on the comparison the expert actually makes. Each mini-batch is built from lists of candidates drawn from a single RNA, and a listwise ListNet objective teaches the network to put the lowest-RMSD candidate of each list first. The output score has no absolute meaning; only its order among candidates of the same sequence matters.
HomRank was developed at A*STAR — the Centre for Frontier AI Research at the Institute of High Performance Computing, the Bioinformatics Institute and the Genome Institute of Singapore — with the National University of Singapore and the University of Zagreb, and posted to bioRxiv in August 2025. Its benchmark and homogeneous-batch training scheme were later adopted by SIRGE, a sequence-conditioned evaluator from the same group.
HomRank keeps the architecture of ARES (Townshend et al., 2021): an atomic embedding layer, three E(3)-equivariant blocks built with e3nn, and three fully connected layers producing the score. Only the sampler and the objective change. Ground-truth RMSDs are converted to reciprocal ranking scores, predicted scores pass through a softmax with temperature 0.1, and the loss is the cross-entropy between the two top-1 distributions. It was trained for one epoch with Adam (learning rate 5e-3) on four A100 GPUs, 32 lists per GPU. A variant, HomRank-PT, first pretrains with ARES's regression loss.
The dataset comprises 190 non-redundant single-chain RNAs: 151 new PDB structures with 1,000 molecular-dynamics candidates each, the 18 ARES training and validation RNAs, and 21 RNA-Puzzles targets with Rosetta FARFAR2 decoys. Clustering yields CLS-0 (146 RNAs), CLS-1 (21) and CLS-2 (23); training uses 131 CLS-0 RNAs after removing its RNA-Puzzles targets, while CLS-1 and CLS-2 become the TestCLS-1 and TestCLS-2 benchmarks. On RNA-Puzzles HomRank puts a near-native candidate (RMSD ≤ 2 Å) first for 95% of targets and within the top five for all of them, against 90% for cgRNASP and HomRank-PT, 86% for ARES+ (ARES retrained on the same data), 71% for ARES and 52% for lociPARSE. On TestCLS-1 its top-1 success is 86%, against 43% for cgRNASP and 33% for ARES. Across predictions from six structure predictors, including trRosettaRNA and AlphaFold 3, for 84 RNAs, it is the only evaluator tested whose Spearman correlation with RMSD is positive (0.227).
HomRank is a selection step for RNA 3D modelling. Sampling methods such as Rosetta FARFAR2 and molecular dynamics, and neural predictors such as RoseTTAFoldNA, produce candidate pools whose best member is not identified by the generator; HomRank scores the pool and returns an ordering, so a modeller can carry the top few conformations into docking, design or comparison with experimental data. The released checkpoint applies to new RNAs without retraining, including pooled outputs from several predictors.
HomRank's argument is that for RNA structure evaluation the training objective and the data split matter as much as the network: with the ARES backbone unchanged, retraining on a clustered dataset alone lifts ARES's RNA-Puzzles top-1 success from 71% to 86%, and the ranking objective takes it to 95%. Performance on the TestCLS-2 cluster remains poor for every method including HomRank, which the authors attribute to the scarcity of solved RNA structures, and they name the ARES backbone as the model's main constraint. The repository holds training and inference code and a trained checkpoint but no license, and the curated molecular-dynamics candidate sets are available from the authors only on request. The preprint has not yet been peer reviewed.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.